TensorFlow中如何简洁实现张量连续重复？将[1,2,3]转为[1,1,2,2,3,3]

Hey there! Your current approach to transform [1,2,3] into [1,1,2,2,3,3] is totally valid, but you’re right—it can be way more concise. Let’s look at the cleanest ways to do this in TensorFlow, especially tailored for your Gaussian Mixture Model (GMM) use case.

The Most Direct Method: tf.repeat() #

The simplest solution is to use TensorFlow’s tf.repeat() function, which is explicitly designed to repeat individual elements of a tensor (unlike tf.tile(), which repeats the entire tensor block). Here’s how it works:

import tensorflow as tf

val = tf.constant([1., 2., 3.])
result = tf.repeat(val, repeats=2)  # Output: [1., 1., 2., 2., 3., 3.]

That’s it—one line to get exactly the tensor you need! This is perfect for GMM scenarios, where you often need to repeat per-component parameters (like means, variances, or weights) across multiple samples or dimensions.

Why Your Original Approach Works (But Isn’t Ideal) #

Just to confirm, your initial code does work with a small syntax fix: tf.reshape(tiled, (2, 3)) instead of tf.reshape(2,3)—but it’s doing unnecessary steps. Let’s correct it for completeness:

val = tf.constant([1.,2.,3.])
tiled = tf.tile(val, [2])  # [1.,2.,3.,1.,2.,3.]
reshaped = tf.reshape(tiled, (2,3))  # [[1.,2.,3.], [1.,2.,3.]]
transposed = tf.transpose(reshaped)  # [[1.,1.], [2.,2.], [3.,3.]]
flattened = tf.reshape(transposed, (6,))  # [1.,1.,2.,2.,3.,3.]

But as you noticed, this is roundabout. tf.repeat() cuts straight to the chase by targeting individual elements for repetition.

Bonus: Batch-Friendly Version #

If you’re working with batch tensors (common in GMM training), tf.repeat() scales seamlessly. For example, if you have a batch of component parameters shaped (batch_size, num_components), you can repeat each component’s value across a new dimension like this:

batch_val = tf.constant([[1.,2.,3.], [4.,5.,6.]])
# Repeat each element 2 times along the last dimension
batch_result = tf.repeat(batch_val, repeats=2, axis=-1)
# Output: [[1.,1.,2.,2.,3.,3.], [4.,4.,5.,5.,6.,6.]]

This is super handy when you need to align component parameters with sample-wise computations in your GMM.

内容的提问来源于stack exchange，提问作者Sam Bobel